1 // SPDX-License-Identifier: GPL-2.0-only
3 * linux/fs/nfs/direct.c
5 * Copyright (C) 2003 by Chuck Lever <cel@netapp.com>
7 * High-performance uncached I/O for the Linux NFS client
9 * There are important applications whose performance or correctness
10 * depends on uncached access to file data. Database clusters
11 * (multiple copies of the same instance running on separate hosts)
12 * implement their own cache coherency protocol that subsumes file
13 * system cache protocols. Applications that process datasets
14 * considerably larger than the client's memory do not always benefit
15 * from a local cache. A streaming video server, for instance, has no
16 * need to cache the contents of a file.
18 * When an application requests uncached I/O, all read and write requests
19 * are made directly to the server; data stored or fetched via these
20 * requests is not cached in the Linux page cache. The client does not
21 * correct unaligned requests from applications. All requested bytes are
22 * held on permanent storage before a direct write system call returns to
25 * Solaris implements an uncached I/O facility called directio() that
26 * is used for backups and sequential I/O to very large files. Solaris
27 * also supports uncaching whole NFS partitions with "-o forcedirectio,"
28 * an undocumented mount option.
30 * Designed by Jeff Kimmel, Chuck Lever, and Trond Myklebust, with
31 * help from Andrew Morton.
33 * 18 Dec 2001 Initial implementation for 2.4 --cel
34 * 08 Jul 2002 Version for 2.4.19, with bug fixes --trondmy
35 * 08 Jun 2003 Port to 2.5 APIs --cel
36 * 31 Mar 2004 Handle direct I/O without VFS support --cel
37 * 15 Sep 2004 Parallel async reads --cel
38 * 04 May 2005 support O_DIRECT with aio --cel
42 #include <linux/errno.h>
43 #include <linux/sched.h>
44 #include <linux/kernel.h>
45 #include <linux/file.h>
46 #include <linux/pagemap.h>
47 #include <linux/kref.h>
48 #include <linux/slab.h>
49 #include <linux/task_io_accounting_ops.h>
50 #include <linux/module.h>
52 #include <linux/nfs_fs.h>
53 #include <linux/nfs_page.h>
54 #include <linux/sunrpc/clnt.h>
56 #include <linux/uaccess.h>
57 #include <linux/atomic.h>
65 #define NFSDBG_FACILITY NFSDBG_VFS
67 static struct kmem_cache *nfs_direct_cachep;
69 static const struct nfs_pgio_completion_ops nfs_direct_write_completion_ops;
70 static const struct nfs_commit_completion_ops nfs_direct_commit_completion_ops;
71 static void nfs_direct_write_complete(struct nfs_direct_req *dreq);
72 static void nfs_direct_write_schedule_work(struct work_struct *work);
74 static inline void get_dreq(struct nfs_direct_req *dreq)
76 atomic_inc(&dreq->io_count);
79 static inline int put_dreq(struct nfs_direct_req *dreq)
81 return atomic_dec_and_test(&dreq->io_count);
85 nfs_direct_handle_truncated(struct nfs_direct_req *dreq,
86 const struct nfs_pgio_header *hdr,
89 if (!(test_bit(NFS_IOHDR_ERROR, &hdr->flags) ||
90 test_bit(NFS_IOHDR_EOF, &hdr->flags)))
92 if (dreq->max_count >= dreq_len) {
93 dreq->max_count = dreq_len;
94 if (dreq->count > dreq_len)
95 dreq->count = dreq_len;
98 if (test_bit(NFS_IOHDR_ERROR, &hdr->flags) && !dreq->error)
99 dreq->error = hdr->error;
103 nfs_direct_count_bytes(struct nfs_direct_req *dreq,
104 const struct nfs_pgio_header *hdr)
106 loff_t hdr_end = hdr->io_start + hdr->good_bytes;
107 ssize_t dreq_len = 0;
109 if (hdr_end > dreq->io_start)
110 dreq_len = hdr_end - dreq->io_start;
112 nfs_direct_handle_truncated(dreq, hdr, dreq_len);
114 if (dreq_len > dreq->max_count)
115 dreq_len = dreq->max_count;
117 if (dreq->count < dreq_len)
118 dreq->count = dreq_len;
121 static void nfs_direct_truncate_request(struct nfs_direct_req *dreq,
122 struct nfs_page *req)
124 loff_t offs = req_offset(req);
125 size_t req_start = (size_t)(offs - dreq->io_start);
127 if (req_start < dreq->max_count)
128 dreq->max_count = req_start;
129 if (req_start < dreq->count)
130 dreq->count = req_start;
134 * nfs_swap_rw - NFS address space operation for swap I/O
135 * @iocb: target I/O control block
138 * Perform IO to the swap-file. This is much like direct IO.
140 int nfs_swap_rw(struct kiocb *iocb, struct iov_iter *iter)
144 VM_BUG_ON(iov_iter_count(iter) != PAGE_SIZE);
146 if (iov_iter_rw(iter) == READ)
147 ret = nfs_file_direct_read(iocb, iter, true);
149 ret = nfs_file_direct_write(iocb, iter, true);
155 static void nfs_direct_release_pages(struct page **pages, unsigned int npages)
158 for (i = 0; i < npages; i++)
162 void nfs_init_cinfo_from_dreq(struct nfs_commit_info *cinfo,
163 struct nfs_direct_req *dreq)
165 cinfo->inode = dreq->inode;
166 cinfo->mds = &dreq->mds_cinfo;
167 cinfo->ds = &dreq->ds_cinfo;
169 cinfo->completion_ops = &nfs_direct_commit_completion_ops;
172 static inline struct nfs_direct_req *nfs_direct_req_alloc(void)
174 struct nfs_direct_req *dreq;
176 dreq = kmem_cache_zalloc(nfs_direct_cachep, GFP_KERNEL);
180 kref_init(&dreq->kref);
181 kref_get(&dreq->kref);
182 init_completion(&dreq->completion);
183 INIT_LIST_HEAD(&dreq->mds_cinfo.list);
184 pnfs_init_ds_commit_info(&dreq->ds_cinfo);
185 INIT_WORK(&dreq->work, nfs_direct_write_schedule_work);
186 spin_lock_init(&dreq->lock);
191 static void nfs_direct_req_free(struct kref *kref)
193 struct nfs_direct_req *dreq = container_of(kref, struct nfs_direct_req, kref);
195 pnfs_release_ds_info(&dreq->ds_cinfo, dreq->inode);
196 if (dreq->l_ctx != NULL)
197 nfs_put_lock_context(dreq->l_ctx);
198 if (dreq->ctx != NULL)
199 put_nfs_open_context(dreq->ctx);
200 kmem_cache_free(nfs_direct_cachep, dreq);
203 static void nfs_direct_req_release(struct nfs_direct_req *dreq)
205 kref_put(&dreq->kref, nfs_direct_req_free);
208 ssize_t nfs_dreq_bytes_left(struct nfs_direct_req *dreq)
210 return dreq->bytes_left;
212 EXPORT_SYMBOL_GPL(nfs_dreq_bytes_left);
215 * Collects and returns the final error value/byte-count.
217 static ssize_t nfs_direct_wait(struct nfs_direct_req *dreq)
219 ssize_t result = -EIOCBQUEUED;
221 /* Async requests don't wait here */
225 result = wait_for_completion_killable(&dreq->completion);
228 result = dreq->count;
229 WARN_ON_ONCE(dreq->count < 0);
232 result = dreq->error;
235 return (ssize_t) result;
239 * Synchronous I/O uses a stack-allocated iocb. Thus we can't trust
240 * the iocb is still valid here if this is a synchronous request.
242 static void nfs_direct_complete(struct nfs_direct_req *dreq)
244 struct inode *inode = dreq->inode;
246 inode_dio_end(inode);
249 long res = (long) dreq->error;
250 if (dreq->count != 0) {
251 res = (long) dreq->count;
252 WARN_ON_ONCE(dreq->count < 0);
254 dreq->iocb->ki_complete(dreq->iocb, res);
257 complete(&dreq->completion);
259 nfs_direct_req_release(dreq);
262 static void nfs_direct_read_completion(struct nfs_pgio_header *hdr)
264 unsigned long bytes = 0;
265 struct nfs_direct_req *dreq = hdr->dreq;
267 spin_lock(&dreq->lock);
268 if (test_bit(NFS_IOHDR_REDO, &hdr->flags)) {
269 spin_unlock(&dreq->lock);
273 nfs_direct_count_bytes(dreq, hdr);
274 spin_unlock(&dreq->lock);
276 while (!list_empty(&hdr->pages)) {
277 struct nfs_page *req = nfs_list_entry(hdr->pages.next);
278 struct page *page = req->wb_page;
280 if (!PageCompound(page) && bytes < hdr->good_bytes &&
281 (dreq->flags == NFS_ODIRECT_SHOULD_DIRTY))
282 set_page_dirty(page);
283 bytes += req->wb_bytes;
284 nfs_list_remove_request(req);
285 nfs_release_request(req);
289 nfs_direct_complete(dreq);
293 static void nfs_read_sync_pgio_error(struct list_head *head, int error)
295 struct nfs_page *req;
297 while (!list_empty(head)) {
298 req = nfs_list_entry(head->next);
299 nfs_list_remove_request(req);
300 nfs_release_request(req);
304 static void nfs_direct_pgio_init(struct nfs_pgio_header *hdr)
309 static const struct nfs_pgio_completion_ops nfs_direct_read_completion_ops = {
310 .error_cleanup = nfs_read_sync_pgio_error,
311 .init_hdr = nfs_direct_pgio_init,
312 .completion = nfs_direct_read_completion,
316 * For each rsize'd chunk of the user's buffer, dispatch an NFS READ
317 * operation. If nfs_readdata_alloc() or get_user_pages() fails,
318 * bail and stop sending more reads. Read length accounting is
319 * handled automatically by nfs_direct_read_result(). Otherwise, if
320 * no requests have been sent, just return an error.
323 static ssize_t nfs_direct_read_schedule_iovec(struct nfs_direct_req *dreq,
324 struct iov_iter *iter,
327 struct nfs_pageio_descriptor desc;
328 struct inode *inode = dreq->inode;
329 ssize_t result = -EINVAL;
330 size_t requested_bytes = 0;
331 size_t rsize = max_t(size_t, NFS_SERVER(inode)->rsize, PAGE_SIZE);
333 nfs_pageio_init_read(&desc, dreq->inode, false,
334 &nfs_direct_read_completion_ops);
337 inode_dio_begin(inode);
339 while (iov_iter_count(iter)) {
340 struct page **pagevec;
345 result = iov_iter_get_pages_alloc2(iter, &pagevec,
351 npages = (result + pgbase + PAGE_SIZE - 1) / PAGE_SIZE;
352 for (i = 0; i < npages; i++) {
353 struct nfs_page *req;
354 unsigned int req_len = min_t(size_t, bytes, PAGE_SIZE - pgbase);
355 /* XXX do we need to do the eof zeroing found in async_filler? */
356 req = nfs_create_request(dreq->ctx, pagevec[i],
359 result = PTR_ERR(req);
362 req->wb_index = pos >> PAGE_SHIFT;
363 req->wb_offset = pos & ~PAGE_MASK;
364 if (!nfs_pageio_add_request(&desc, req)) {
365 result = desc.pg_error;
366 nfs_release_request(req);
371 requested_bytes += req_len;
373 dreq->bytes_left -= req_len;
375 nfs_direct_release_pages(pagevec, npages);
381 nfs_pageio_complete(&desc);
384 * If no bytes were started, return the error, and let the
385 * generic layer handle the completion.
387 if (requested_bytes == 0) {
388 inode_dio_end(inode);
389 nfs_direct_req_release(dreq);
390 return result < 0 ? result : -EIO;
394 nfs_direct_complete(dreq);
395 return requested_bytes;
399 * nfs_file_direct_read - file direct read operation for NFS files
400 * @iocb: target I/O control block
401 * @iter: vector of user buffers into which to read data
402 * @swap: flag indicating this is swap IO, not O_DIRECT IO
404 * We use this function for direct reads instead of calling
405 * generic_file_aio_read() in order to avoid gfar's check to see if
406 * the request starts before the end of the file. For that check
407 * to work, we must generate a GETATTR before each direct read, and
408 * even then there is a window between the GETATTR and the subsequent
409 * READ where the file size could change. Our preference is simply
410 * to do all reads the application wants, and the server will take
411 * care of managing the end of file boundary.
413 * This function also eliminates unnecessarily updating the file's
414 * atime locally, as the NFS server sets the file's atime, and this
415 * client must read the updated atime from the server back into its
418 ssize_t nfs_file_direct_read(struct kiocb *iocb, struct iov_iter *iter,
421 struct file *file = iocb->ki_filp;
422 struct address_space *mapping = file->f_mapping;
423 struct inode *inode = mapping->host;
424 struct nfs_direct_req *dreq;
425 struct nfs_lock_context *l_ctx;
426 ssize_t result, requested;
427 size_t count = iov_iter_count(iter);
428 nfs_add_stats(mapping->host, NFSIOS_DIRECTREADBYTES, count);
430 dfprintk(FILE, "NFS: direct read(%pD2, %zd@%Ld)\n",
431 file, count, (long long) iocb->ki_pos);
437 task_io_account_read(count);
440 dreq = nfs_direct_req_alloc();
445 dreq->bytes_left = dreq->max_count = count;
446 dreq->io_start = iocb->ki_pos;
447 dreq->ctx = get_nfs_open_context(nfs_file_open_context(iocb->ki_filp));
448 l_ctx = nfs_get_lock_context(dreq->ctx);
450 result = PTR_ERR(l_ctx);
451 nfs_direct_req_release(dreq);
455 if (!is_sync_kiocb(iocb))
458 if (user_backed_iter(iter))
459 dreq->flags = NFS_ODIRECT_SHOULD_DIRTY;
462 nfs_start_io_direct(inode);
464 NFS_I(inode)->read_io += count;
465 requested = nfs_direct_read_schedule_iovec(dreq, iter, iocb->ki_pos);
468 nfs_end_io_direct(inode);
471 result = nfs_direct_wait(dreq);
474 iocb->ki_pos += result;
476 iov_iter_revert(iter, requested);
482 nfs_direct_req_release(dreq);
487 static void nfs_direct_add_page_head(struct list_head *list,
488 struct nfs_page *req)
490 struct nfs_page *head = req->wb_head;
492 if (!list_empty(&head->wb_list) || !nfs_lock_request(head))
494 if (!list_empty(&head->wb_list)) {
495 nfs_unlock_request(head);
498 list_add(&head->wb_list, list);
499 kref_get(&head->wb_kref);
500 kref_get(&head->wb_kref);
503 static void nfs_direct_join_group(struct list_head *list, struct inode *inode)
505 struct nfs_page *req, *subreq;
507 list_for_each_entry(req, list, wb_list) {
508 if (req->wb_head != req) {
509 nfs_direct_add_page_head(&req->wb_list, req);
512 subreq = req->wb_this_page;
517 * Remove subrequests from this list before freeing
518 * them in the call to nfs_join_page_group().
520 if (!list_empty(&subreq->wb_list)) {
521 nfs_list_remove_request(subreq);
522 nfs_release_request(subreq);
524 } while ((subreq = subreq->wb_this_page) != req);
525 nfs_join_page_group(req, inode);
530 nfs_direct_write_scan_commit_list(struct inode *inode,
531 struct list_head *list,
532 struct nfs_commit_info *cinfo)
534 mutex_lock(&NFS_I(cinfo->inode)->commit_mutex);
535 pnfs_recover_commit_reqs(list, cinfo);
536 nfs_scan_commit_list(&cinfo->mds->list, list, cinfo, 0);
537 mutex_unlock(&NFS_I(cinfo->inode)->commit_mutex);
540 static void nfs_direct_write_reschedule(struct nfs_direct_req *dreq)
542 struct nfs_pageio_descriptor desc;
543 struct nfs_page *req;
545 struct nfs_commit_info cinfo;
547 nfs_init_cinfo_from_dreq(&cinfo, dreq);
548 nfs_direct_write_scan_commit_list(dreq->inode, &reqs, &cinfo);
550 nfs_direct_join_group(&reqs, dreq->inode);
552 nfs_clear_pnfs_ds_commit_verifiers(&dreq->ds_cinfo);
555 nfs_pageio_init_write(&desc, dreq->inode, FLUSH_STABLE, false,
556 &nfs_direct_write_completion_ops);
559 while (!list_empty(&reqs)) {
560 req = nfs_list_entry(reqs.next);
561 /* Bump the transmission count */
563 if (!nfs_pageio_add_request(&desc, req)) {
564 spin_lock(&dreq->lock);
565 if (dreq->error < 0) {
566 desc.pg_error = dreq->error;
567 } else if (desc.pg_error != -EAGAIN) {
570 desc.pg_error = -EIO;
571 dreq->error = desc.pg_error;
573 dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
574 spin_unlock(&dreq->lock);
577 nfs_release_request(req);
579 nfs_pageio_complete(&desc);
581 while (!list_empty(&reqs)) {
582 req = nfs_list_entry(reqs.next);
583 nfs_list_remove_request(req);
584 nfs_unlock_and_release_request(req);
585 if (desc.pg_error == -EAGAIN) {
586 nfs_mark_request_commit(req, NULL, &cinfo, 0);
588 spin_lock(&dreq->lock);
589 nfs_direct_truncate_request(dreq, req);
590 spin_unlock(&dreq->lock);
591 nfs_release_request(req);
596 nfs_direct_write_complete(dreq);
599 static void nfs_direct_commit_complete(struct nfs_commit_data *data)
601 const struct nfs_writeverf *verf = data->res.verf;
602 struct nfs_direct_req *dreq = data->dreq;
603 struct nfs_commit_info cinfo;
604 struct nfs_page *req;
605 int status = data->task.tk_status;
607 trace_nfs_direct_commit_complete(dreq);
610 /* Errors in commit are fatal */
611 dreq->error = status;
612 dreq->flags = NFS_ODIRECT_DONE;
614 status = dreq->error;
617 nfs_init_cinfo_from_dreq(&cinfo, dreq);
619 while (!list_empty(&data->pages)) {
620 req = nfs_list_entry(data->pages.next);
621 nfs_list_remove_request(req);
623 spin_lock(&dreq->lock);
624 nfs_direct_truncate_request(dreq, req);
625 spin_unlock(&dreq->lock);
626 nfs_release_request(req);
627 } else if (!nfs_write_match_verf(verf, req)) {
628 dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
630 * Despite the reboot, the write was successful,
634 nfs_mark_request_commit(req, NULL, &cinfo, 0);
636 nfs_release_request(req);
637 nfs_unlock_and_release_request(req);
640 if (nfs_commit_end(cinfo.mds))
641 nfs_direct_write_complete(dreq);
644 static void nfs_direct_resched_write(struct nfs_commit_info *cinfo,
645 struct nfs_page *req)
647 struct nfs_direct_req *dreq = cinfo->dreq;
649 trace_nfs_direct_resched_write(dreq);
651 spin_lock(&dreq->lock);
652 if (dreq->flags != NFS_ODIRECT_DONE)
653 dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
654 spin_unlock(&dreq->lock);
655 nfs_mark_request_commit(req, NULL, cinfo, 0);
658 static const struct nfs_commit_completion_ops nfs_direct_commit_completion_ops = {
659 .completion = nfs_direct_commit_complete,
660 .resched_write = nfs_direct_resched_write,
663 static void nfs_direct_commit_schedule(struct nfs_direct_req *dreq)
666 struct nfs_commit_info cinfo;
669 nfs_init_cinfo_from_dreq(&cinfo, dreq);
670 nfs_scan_commit(dreq->inode, &mds_list, &cinfo);
671 res = nfs_generic_commit_list(dreq->inode, &mds_list, 0, &cinfo);
672 if (res < 0) /* res == -ENOMEM */
673 nfs_direct_write_reschedule(dreq);
676 static void nfs_direct_write_clear_reqs(struct nfs_direct_req *dreq)
678 struct nfs_commit_info cinfo;
679 struct nfs_page *req;
682 nfs_init_cinfo_from_dreq(&cinfo, dreq);
683 nfs_direct_write_scan_commit_list(dreq->inode, &reqs, &cinfo);
685 while (!list_empty(&reqs)) {
686 req = nfs_list_entry(reqs.next);
687 nfs_list_remove_request(req);
688 nfs_direct_truncate_request(dreq, req);
689 nfs_release_request(req);
690 nfs_unlock_and_release_request(req);
694 static void nfs_direct_write_schedule_work(struct work_struct *work)
696 struct nfs_direct_req *dreq = container_of(work, struct nfs_direct_req, work);
697 int flags = dreq->flags;
701 case NFS_ODIRECT_DO_COMMIT:
702 nfs_direct_commit_schedule(dreq);
704 case NFS_ODIRECT_RESCHED_WRITES:
705 nfs_direct_write_reschedule(dreq);
708 nfs_direct_write_clear_reqs(dreq);
709 nfs_zap_mapping(dreq->inode, dreq->inode->i_mapping);
710 nfs_direct_complete(dreq);
714 static void nfs_direct_write_complete(struct nfs_direct_req *dreq)
716 trace_nfs_direct_write_complete(dreq);
717 queue_work(nfsiod_workqueue, &dreq->work); /* Calls nfs_direct_write_schedule_work */
720 static void nfs_direct_write_completion(struct nfs_pgio_header *hdr)
722 struct nfs_direct_req *dreq = hdr->dreq;
723 struct nfs_commit_info cinfo;
724 struct nfs_page *req = nfs_list_entry(hdr->pages.next);
725 int flags = NFS_ODIRECT_DONE;
727 trace_nfs_direct_write_completion(dreq);
729 nfs_init_cinfo_from_dreq(&cinfo, dreq);
731 spin_lock(&dreq->lock);
732 if (test_bit(NFS_IOHDR_REDO, &hdr->flags)) {
733 spin_unlock(&dreq->lock);
737 nfs_direct_count_bytes(dreq, hdr);
738 if (test_bit(NFS_IOHDR_UNSTABLE_WRITES, &hdr->flags) &&
739 !test_bit(NFS_IOHDR_ERROR, &hdr->flags)) {
741 dreq->flags = NFS_ODIRECT_DO_COMMIT;
744 spin_unlock(&dreq->lock);
746 while (!list_empty(&hdr->pages)) {
748 req = nfs_list_entry(hdr->pages.next);
749 nfs_list_remove_request(req);
750 if (flags == NFS_ODIRECT_DO_COMMIT) {
751 kref_get(&req->wb_kref);
752 memcpy(&req->wb_verf, &hdr->verf.verifier,
753 sizeof(req->wb_verf));
754 nfs_mark_request_commit(req, hdr->lseg, &cinfo,
756 } else if (flags == NFS_ODIRECT_RESCHED_WRITES) {
757 kref_get(&req->wb_kref);
758 nfs_mark_request_commit(req, NULL, &cinfo, 0);
760 nfs_unlock_and_release_request(req);
765 nfs_direct_write_complete(dreq);
769 static void nfs_write_sync_pgio_error(struct list_head *head, int error)
771 struct nfs_page *req;
773 while (!list_empty(head)) {
774 req = nfs_list_entry(head->next);
775 nfs_list_remove_request(req);
776 nfs_unlock_and_release_request(req);
780 static void nfs_direct_write_reschedule_io(struct nfs_pgio_header *hdr)
782 struct nfs_direct_req *dreq = hdr->dreq;
784 trace_nfs_direct_write_reschedule_io(dreq);
786 spin_lock(&dreq->lock);
787 if (dreq->error == 0) {
788 dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
789 /* fake unstable write to let common nfs resend pages */
790 hdr->verf.committed = NFS_UNSTABLE;
791 hdr->good_bytes = hdr->args.offset + hdr->args.count -
794 spin_unlock(&dreq->lock);
797 static const struct nfs_pgio_completion_ops nfs_direct_write_completion_ops = {
798 .error_cleanup = nfs_write_sync_pgio_error,
799 .init_hdr = nfs_direct_pgio_init,
800 .completion = nfs_direct_write_completion,
801 .reschedule_io = nfs_direct_write_reschedule_io,
806 * NB: Return the value of the first error return code. Subsequent
807 * errors after the first one are ignored.
810 * For each wsize'd chunk of the user's buffer, dispatch an NFS WRITE
811 * operation. If nfs_writedata_alloc() or get_user_pages() fails,
812 * bail and stop sending more writes. Write length accounting is
813 * handled automatically by nfs_direct_write_result(). Otherwise, if
814 * no requests have been sent, just return an error.
816 static ssize_t nfs_direct_write_schedule_iovec(struct nfs_direct_req *dreq,
817 struct iov_iter *iter,
818 loff_t pos, int ioflags)
820 struct nfs_pageio_descriptor desc;
821 struct inode *inode = dreq->inode;
822 struct nfs_commit_info cinfo;
824 size_t requested_bytes = 0;
825 size_t wsize = max_t(size_t, NFS_SERVER(inode)->wsize, PAGE_SIZE);
828 trace_nfs_direct_write_schedule_iovec(dreq);
830 nfs_pageio_init_write(&desc, inode, ioflags, false,
831 &nfs_direct_write_completion_ops);
834 inode_dio_begin(inode);
836 NFS_I(inode)->write_io += iov_iter_count(iter);
837 while (iov_iter_count(iter)) {
838 struct page **pagevec;
843 result = iov_iter_get_pages_alloc2(iter, &pagevec,
849 npages = (result + pgbase + PAGE_SIZE - 1) / PAGE_SIZE;
850 for (i = 0; i < npages; i++) {
851 struct nfs_page *req;
852 unsigned int req_len = min_t(size_t, bytes, PAGE_SIZE - pgbase);
854 req = nfs_create_request(dreq->ctx, pagevec[i],
857 result = PTR_ERR(req);
861 if (desc.pg_error < 0) {
862 nfs_free_request(req);
863 result = desc.pg_error;
869 requested_bytes += req_len;
871 dreq->bytes_left -= req_len;
874 nfs_mark_request_commit(req, NULL, &cinfo, 0);
878 nfs_lock_request(req);
879 req->wb_index = pos >> PAGE_SHIFT;
880 req->wb_offset = pos & ~PAGE_MASK;
881 if (nfs_pageio_add_request(&desc, req))
884 /* Exit on hard errors */
885 if (desc.pg_error < 0 && desc.pg_error != -EAGAIN) {
886 result = desc.pg_error;
887 nfs_unlock_and_release_request(req);
891 /* If the error is soft, defer remaining requests */
892 nfs_init_cinfo_from_dreq(&cinfo, dreq);
893 spin_lock(&dreq->lock);
894 dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
895 spin_unlock(&dreq->lock);
896 nfs_unlock_request(req);
897 nfs_mark_request_commit(req, NULL, &cinfo, 0);
901 nfs_direct_release_pages(pagevec, npages);
906 nfs_pageio_complete(&desc);
909 * If no bytes were started, return the error, and let the
910 * generic layer handle the completion.
912 if (requested_bytes == 0) {
913 inode_dio_end(inode);
914 nfs_direct_req_release(dreq);
915 return result < 0 ? result : -EIO;
919 nfs_direct_write_complete(dreq);
920 return requested_bytes;
924 * nfs_file_direct_write - file direct write operation for NFS files
925 * @iocb: target I/O control block
926 * @iter: vector of user buffers from which to write data
927 * @swap: flag indicating this is swap IO, not O_DIRECT IO
929 * We use this function for direct writes instead of calling
930 * generic_file_aio_write() in order to avoid taking the inode
931 * semaphore and updating the i_size. The NFS server will set
932 * the new i_size and this client must read the updated size
933 * back into its cache. We let the server do generic write
934 * parameter checking and report problems.
936 * We eliminate local atime updates, see direct read above.
938 * We avoid unnecessary page cache invalidations for normal cached
939 * readers of this file.
941 * Note that O_APPEND is not supported for NFS direct writes, as there
942 * is no atomic O_APPEND write facility in the NFS protocol.
944 ssize_t nfs_file_direct_write(struct kiocb *iocb, struct iov_iter *iter,
947 ssize_t result, requested;
949 struct file *file = iocb->ki_filp;
950 struct address_space *mapping = file->f_mapping;
951 struct inode *inode = mapping->host;
952 struct nfs_direct_req *dreq;
953 struct nfs_lock_context *l_ctx;
956 dfprintk(FILE, "NFS: direct write(%pD2, %zd@%Ld)\n",
957 file, iov_iter_count(iter), (long long) iocb->ki_pos);
960 /* bypass generic checks */
961 result = iov_iter_count(iter);
963 result = generic_write_checks(iocb, iter);
967 nfs_add_stats(mapping->host, NFSIOS_DIRECTWRITTENBYTES, count);
970 end = (pos + iov_iter_count(iter) - 1) >> PAGE_SHIFT;
972 task_io_account_write(count);
975 dreq = nfs_direct_req_alloc();
980 dreq->bytes_left = dreq->max_count = count;
981 dreq->io_start = pos;
982 dreq->ctx = get_nfs_open_context(nfs_file_open_context(iocb->ki_filp));
983 l_ctx = nfs_get_lock_context(dreq->ctx);
985 result = PTR_ERR(l_ctx);
986 nfs_direct_req_release(dreq);
990 if (!is_sync_kiocb(iocb))
992 pnfs_init_ds_commit_info_ops(&dreq->ds_cinfo, inode);
995 requested = nfs_direct_write_schedule_iovec(dreq, iter, pos,
998 nfs_start_io_direct(inode);
1000 requested = nfs_direct_write_schedule_iovec(dreq, iter, pos,
1003 if (mapping->nrpages) {
1004 invalidate_inode_pages2_range(mapping,
1005 pos >> PAGE_SHIFT, end);
1008 nfs_end_io_direct(inode);
1011 if (requested > 0) {
1012 result = nfs_direct_wait(dreq);
1014 requested -= result;
1015 iocb->ki_pos = pos + result;
1016 /* XXX: should check the generic_write_sync retval */
1017 generic_write_sync(iocb, result);
1019 iov_iter_revert(iter, requested);
1023 nfs_fscache_invalidate(inode, FSCACHE_INVAL_DIO_WRITE);
1025 nfs_direct_req_release(dreq);
1031 * nfs_init_directcache - create a slab cache for nfs_direct_req structures
1034 int __init nfs_init_directcache(void)
1036 nfs_direct_cachep = kmem_cache_create("nfs_direct_cache",
1037 sizeof(struct nfs_direct_req),
1038 0, (SLAB_RECLAIM_ACCOUNT|
1041 if (nfs_direct_cachep == NULL)
1048 * nfs_destroy_directcache - destroy the slab cache for nfs_direct_req structures
1051 void nfs_destroy_directcache(void)
1053 kmem_cache_destroy(nfs_direct_cachep);